You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Parity #478] Complete learned-model-state diagnostics for every canonical estimator family #492
Performance must not outrun semantic evidence. Current benchmark infrastructure has strong state diagnostics for only a subset of rows (notably PCA/KMeans and selected later additions); score parity alone is insufficient for optimizer work.
Add estimator-state comparisons for every canonical family: linear coefficients/intercepts and objective/convergence state; SVM weights/support vectors/dual coefficients/intercepts; tree node structure/thresholds/features/leaves; forest per-tree/bootstrap/RNG structure; GaussianNB priors/means/variances; KMeans centers/inertia/iterations; PCA means/components/singular values; Ridge/Lasso/LinearRegression coefficients/intercepts/objective state; KernelRidge dual coefficients/kernel parameters.
Acceptance: all 19 canonical rows report state_coverage=covered or an explicit semantically justified exception; state deltas are frozen in disparity history; performance classification is blocked when required state evidence regresses beyond contract.
I worked through a semantic hole in learned-state parity: raw coordinate comparison is not valid for estimator families with exact state symmetries. KMeans centers must be compared modulo cluster-label permutation; PCA retained state should include an invariant subspace/projector comparison because sign flips and, under repeated/degenerate singular values, orthogonal rotations of the retained basis can represent the same model. A general contract is orbit distance d_G(theta,theta') = inf_g d(theta,g theta') for the estimator's semantics-preserving symmetry group. Concrete counterexamples: three KMeans centers can have raw max error 10 yet permutation-aware error 0; two PCA component matrices can have raw max error 1 yet identical V^T V projectors. I added the full derivation and an idiomatic Flow witness to godofecht/OpenProjects under ml/symmetry-aware-learned-state-parity.{md,flow}. Recommendation for #492: keep model_state_coverage.py as inventory, but require estimator-specific semantic-state schemas: KMeans minimum-cost center matching before state deltas; PCA mean+spectrum+projector parity, with sign-aligned row comparison only for nondegenerate components. Performance eligibility should depend on those semantic certificates, not merely the presence of any diagnostic key.
Activity
godofecht commentedon Aug 26, 2026
I worked through a semantic hole in learned-state parity: raw coordinate comparison is not valid for estimator families with exact state symmetries. KMeans centers must be compared modulo cluster-label permutation; PCA retained state should include an invariant subspace/projector comparison because sign flips and, under repeated/degenerate singular values, orthogonal rotations of the retained basis can represent the same model. A general contract is orbit distance d_G(theta,theta') = inf_g d(theta,g theta') for the estimator's semantics-preserving symmetry group. Concrete counterexamples: three KMeans centers can have raw max error 10 yet permutation-aware error 0; two PCA component matrices can have raw max error 1 yet identical V^T V projectors. I added the full derivation and an idiomatic Flow witness to godofecht/OpenProjects under ml/symmetry-aware-learned-state-parity.{md,flow}. Recommendation for #492: keep model_state_coverage.py as inventory, but require estimator-specific semantic-state schemas: KMeans minimum-cost center matching before state deltas; PCA mean+spectrum+projector parity, with sign-aligned row comparison only for nondegenerate components. Performance eligibility should depend on those semantic certificates, not merely the presence of any diagnostic key.